"""Generate the next counselor response with the merged CounselTrajLM model.""" import argparse import json from pathlib import Path import torch from transformers import AutoModelForCausalLM, AutoTokenizer SYSTEM_PROMPT = ( "你是一名专业、温和、自然的线上心理咨询师。" "请根据对话历史生成下一句咨询师回复。" "回复只写咨询师要说的话,不要写说话人名称,不要写对话标签。" ) def load_messages(path): messages = json.loads(Path(path).read_text(encoding="utf-8")) if not isinstance(messages, list) or not messages: raise ValueError("Provide a non-empty JSON list of chat messages.") for i, message in enumerate(messages): if not isinstance(message, dict): raise ValueError("Each message must be a JSON object.") if message.get("role") not in {"system", "user", "assistant"}: raise ValueError("Message roles must be system, user, or assistant.") if message["role"] == "system" and i != 0: raise ValueError("A system message may appear only at the beginning.") if not isinstance(message.get("content"), str) or not message["content"].strip(): raise ValueError("Each message must have non-empty text content.") if messages[-1]["role"] != "user": raise ValueError("The dialogue must end with a client (user) message.") if messages[0]["role"] != "system": messages.insert(0, {"role": "system", "content": SYSTEM_PROMPT}) return messages def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--model", default="UEC-InabaLab/CounselTrajLM-Qwen2.5-7B") parser.add_argument("--messages", type=Path, default=Path(__file__).parent / "examples/history.json") parser.add_argument("--max-new-tokens", type=int, default=160) parser.add_argument("--greedy", action="store_true") args = parser.parse_args() messages = load_messages(args.messages) tokenizer = AutoTokenizer.from_pretrained(args.model) model = AutoModelForCausalLM.from_pretrained( args.model, dtype=torch.bfloat16, device_map="auto" ).eval() prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = tokenizer(prompt, return_tensors="pt").to(model.device) if inputs.input_ids.shape[-1] + args.max_new_tokens > model.config.max_position_embeddings: raise ValueError("The input plus requested output exceeds the model context window.") generation = { "max_new_tokens": args.max_new_tokens, "do_sample": not args.greedy, "repetition_penalty": 1.05, } if not args.greedy: generation.update(temperature=0.7, top_p=0.8, top_k=20) with torch.inference_mode(): outputs = model.generate(**inputs, **generation) reply = tokenizer.decode(outputs[0, inputs.input_ids.shape[-1]:], skip_special_tokens=True) print(reply.strip()) if __name__ == "__main__": main()